selects-onnx

ONNX Runtime weights for the selects desktop app (photo culling / curation). Replaces the PyTorch + CUDA runtime with ONNX Runtime (DirectML on Windows).

File	Model	Precision	Task
siglip_vision.onnx	SigLIP SO400M vision tower	fp16	image embeddings
siglip_text.onnx	SigLIP SO400M text tower	fp16	text embeddings
nafnet.onnx	NAFNet GoPro width32	fp16	deblur
ram_plus.onnx (+ .data)	RAM++ (Recognize Anything Plus)	fp32	tagging
csrnet.onnx (+ .data)	CSRNet FiveK	fp32	retouch
zero_dce.onnx (+ .data)	Zero-DCE++	fp32	low-light
ram_meta.npz, ram_tags.json	RAM++ thresholds / tag vocab	-	post-processing

All exports are parity-checked against the original PyTorch models. RAM++ is kept fp32 (fp16 conversion trips a Swin shifted-window mask cast; ~400MB not worth it). fp16 models verified cos >= 0.9998.